Research On API Recommendation Model Based On Siamese Network And Graph Convolution

In the modern era.Software development is becoming more and more complex,the importance of API(Application Programming Interface)as an indispensable part of software development is becoming prominent.Existing API recommendation models focus more on matching between texts,lack of research on the spatial structure of APIs,and it is difficult to solve the problems of knowledge gaps and representation differences in the API recommendation field?Therefore,based on the problems of the existing API recommendation model,this article constructs a knowledge graph about API,and proposes a new API recommendation model combined with Siamese Network and Graph Convolution Network,which improves the quality and accuracy of API recommendation.The main work is as follows:1.Build an API knowledge graph.A heterogeneous knowledge graph is proposed,which extracts knowledge from documents and third-party knowledge communities respectively,filters the redundant knowledge among them,and aligns and reconstructs the filtered knowledge,so that these heterogeneous knowledges constitute An API knowledge graph with more than 300,000 nodes.2.An API recommendation model based on Siamese Network and BERT is proposed.Obtain the natural language question vector representation based on BERT,extract API knowledge from the question and introduce the Siamese network structure base on these extracted knowledges to perform semantic correction on the sentence vector to make it contain richer API information and fill the BERT pre-trained model has defects such as lack of API knowledge and difficulty in fine tuning.3.Research on the embedding of API knowledge by Graph Convolutional Networks.Aiming at the lack of semantics of API knowledge space in existing recommendation models,this paper proposes an API knowledge embedding model based on Graph Convolution Network,which uses the correlation between knowledge in a two-dimensional space to embed API knowledge and adds more High-dimensional spatial structure information.Under the combination of semantic embedding and spatial embedding,the problems of knowledge gap and representation difference in API recommendation system are solved.And based on the BIKER model data,the three indicators of TOP@5,MRR@5,and MAP@5 have reached 76.96%,62.59%,60.59%,respectively,which are all improved on the basis of the BIKER and DeepAPI models.